Benchmarking Yolov8n, Yolov11n, and Yolov12n for Smart Road Pothole Detection

Abstract

Road potholes present a considerable threat to traffic safety, highlighting the need for precise and efficient automatic detection in road-monitoring systems. Although recent research has utilized various YOLO-based models for detecting potholes, there is a scarcity of comparative evaluations of the newest lightweight YOLO architectures under uniform experimental conditions. This study seeks to evaluate the performance of YOLOv8n, YOLOv11n, and YOLOv12n in the context of smart road pothole detection. A quantitative experimental approach was applied using the road pole dataset. Each model underwent training using identical preprocessing, training configurations, and evaluation protocols. Performance metrics included Precision, Recall, mAP50, mAP50–95, inference time, and computational efficiency. The experimental findings indicate that YOLOv11n delivered the best overall results, achieving 90.02% precision, 90.25% mAP50, and 48.35% mAP50–95, along with the quickest inference time of 3.13 ms and the highest end-to-end processing speed. Although YOLOv12n recorded the highest recall at 86.11%, its overall detection performance and computational efficiency were inferior to those of YOLOv11n. These results suggest that YOLOv11n offers the most balanced compromise between detection accuracy and computational efficiency, providing empirical support for choosing lightweight object detection models for intelligent road-monitoring systems.